Why the seemingly repetitive code in DeepSpeedDataLoader?

[litaochenbit]

I was going through the code of DeepSpeedDataLoader and noticed that the DataLoader constuctor are called seperately to create a DataLoader object determined by if statements.

https://github.com/microsoft/DeepSpeed/blob/3324efd9306bc19b3cd7a669b339f84983d3ebf8/deepspeed/runtime/dataloader.py#L126-L159

Why not just:

def _create_dataloader(self):
        self.dataloader = DataLoader(
            self.dataset,
            batch_size=self.batch_size,
            pin_memory=self.pin_memory,
            sampler=self.data_sampler,
            collate_fn=self.collate_fn,  # self.collate_fn is default to be None anyways
            num_workers=self.num_local_io_workers,
            drop_last=self.dataloader_drop_last if self.curriculum_learning_enabled else False  # drop_last is default to be False
        )
        if self.curriculum_learning_enabled:
            self.data_iterator = iter(self.dataloader)
            return self.dataloader
        else:
            self.data = (x for x in self.dataloader)
        return self.dataloader

The code would be much more readable.

So as a new learner, I'm just curious. What is the purpose of the seemingly repetitive code? Is it to make sure that the default values of params in torch.utils.data.DataLoader would not be messed up in case they are changed by PyTorch team?